A Study of Residue Correlation within Protein Sequences and Its Application to Sequence Classification

Christopher Michael Hemmerich, Sun KimEURASIP J. Bioinform. Syst. Biol 2007 · DOI

TL;DR

This paper introduces the Mutual Information Vector (MIV), a novel method for capturing long-range correlations in protein sequences beyond simple adjacent residue interactions. MIV significantly outperforms traditional Mutual Information (MI) methods in classifying protein families, achieving accuracy levels that allow classification without sequence alignment.

What Problem Does This Paper Address?

Proteins are complex molecules whose function is dictated by their amino acid sequence. While sequences are often treated as strings of characters, understanding the intricate relationships between amino acids, especially those far apart in the sequence, is crucial for predicting protein structure and function. Traditional methods for analyzing these relationships, like Mutual Information (MI), primarily focus on adjacent residues and can be sensitive to sampling errors, limiting their effectiveness for classification tasks. This paper aims to develop a more robust method to capture these correlations and demonstrate its utility in classifying protein families without relying on sequence alignment, a computationally intensive process.

Methods

The core innovation is the Mutual Information Vector (MIV). Instead of a single MI score, MIV is a vector where each element represents the MI between amino acid pairs separated by a specific distance (d). This allows for the capture of long-range interactions. The authors explore two primary "alphabets" for representing amino acids:

  • ΣA (Amino Acid Alphabet): The standard 20 amino acids.
  • ΣH (Hydropathy Alphabet): A simplified alphabet of two symbols (hydrophobic/hydrophilic), based on physical properties.

They also investigate different methods for handling gap symbols in sequence alignments (strict, literal, and hybrid) and explore normalization techniques (using joint entropy) to mitigate sampling errors. The MIVs are then used as feature vectors for a nearest neighbor classifier to perform protein family classification.

Datasets

The primary dataset used is the Pfam database, a large collection of protein families. For experiments, a subset of Pfam-A was selected, filtering for families with at least 10 sequences and lengths between 100 and 200 residues. For significance testing, sequences were compared against randomly shuffled sequences generated using the HMMER package's shuffle command to preserve sequence length and residue composition.

Key Insights and Ideas

  • Long-range correlations matter: MIV, by considering residue pairs at various distances, captures more information about protein sequences than traditional MI focused only on adjacent pairs.
  • Hydropathy is powerful: The simplified hydropathy alphabet (ΣH) surprisingly yields better classification results than the full amino acid alphabet (ΣA) when used in MIVs, suggesting that physical properties play a significant role in protein structure and function.
  • Gaps are informative: Treating gap symbols literally or using a hybrid approach in MIV calculation leads to better classification performance than strictly removing them. This highlights that gaps in alignments carry information about sequence relationships.
  • MIV enables alignment-free classification: The MIV approach achieves high accuracy in protein family classification without requiring sequence alignment, a significant computational advantage.

Key Results

  • Significance of MI: For real protein sequences, MI scores (especially MI(0)) are statistically significant compared to random sequences, indicating genuine correlations.
  • MIV outperforms MI: In protein family classification, MIV-based methods achieved significantly higher accuracy than traditional MI(0) scores. The best MIV method reached an accuracy of 85.14%, while the best MI(0) method achieved only 26.73%.
  • Best performing methods: The top-performing MIV methods consistently utilized the hydropathy alphabet (ΣH) and often employed a hybrid gap interpretation.
  • Concatenated MIVs excel: Combining different MIV scoring methods (e.g., ΣH with hybrid gaps and ΣA with a Pfam prior) further boosted classification accuracy, with the best combination reaching 90.99%.

Limitations and Future Work

  • Interpretation of MIV: The authors acknowledge that interpreting the specific meaning of each element in a MIV vector remains an open question.
  • Secondary structure correlation: Further investigation into how MIV relates to protein secondary structures is suggested.
  • Normalized MIV vs. Classification: While normalized MIV scores showed improved information content in some analyses, they didn't always translate to better classification accuracy compared to non-normalized scores, warranting further study.
  • Alphabet size effects: The paper touches upon the effect of alphabet size but suggests deeper exploration into how reducing alphabet size (like to ΣH) impacts entropy and MI estimation, particularly concerning finite sample effects.

Summary updated

BibTeX
@article{hemmerich2007study,
  author = {Christopher Michael Hemmerich and Sun Kim},
  title = {{A Study of Residue Correlation within Protein Sequences and Its Application to Sequence Classification}},
  journal = {EURASIP J. Bioinform. Syst. Biol.},
  year = {2007},
  doi = {10.1155/2007/87356},
}

Loading related content...

Figure